OpenAI agents now finish 46% of mid-length research tasks alone

AI agents illustration: a cursor moving along a chain of task cards

OpenAI published its second "research acceleration" report on November 18, subtitled "The jagged frontier," using internal data to measure how coding agents affect AI R&D. The headline figure: for research tasks that would take a human roughly 2 to 4 hours, agents completed 46% independently in October, versus 23% in March. For tasks of 4 to 8 hours, the share rose over the same period from 13% to 35%.

All of these are OpenAI's own estimates. How "independently completed" is determined is not explained in the media accounts of the report, so it cannot currently be verified from public material.

Same research work, very different results

The subtitle refers to the gaps between tasks. In the window from August 1 to October 18, agents succeeded 79% of the time on tasks such as monitoring a model-training run, 35% on launching training jobs, and only 25% on writing inference and serving-stack code.

OpenAI's conclusion is restrained:

"The agents are not taking over AI research. They are taking over specific slices."

One researcher, a dozen or more agents

Parallel use is deepening. According to the report, more than two-thirds of researchers have at least 4 agents running at once at their daily peak, and more than 10% peak at 16 or more. By research area, the number of recorded researcher work-weeks in safety and alignment grew 77% from January to September, the largest increase of any area.

The report also includes a survey of 101 researchers, who were asked how much they would give up to keep access, for three months, to agents stronger than GPT-5.2. The median answer was 30% of their team's compute, or 40% of its headcount. This is a hypothetical preference survey, and there is currently no sign that OpenAI has adjusted hiring because of it.

$2,400 per person per day

The report converts agent usage into dollars at API list prices. In October, the median researcher's daily usage came to about $2,400, up from $1,500 in September; the 90th percentile rose from $11,000 to $23,000.

A rough calculation, assuming about 250 working days a year: the median researcher's annual agent usage works out to roughly $600,000, and the 90th percentile to about $5.75 million. That is the external list-price basis. OpenAI's actual internal compute cost is far lower, but how much lower has not been disclosed, so the figures show only the scale of usage and cannot be read as real spending.

Read against the previous report

The September report framed it differently: research agents reached 3.1 workdays of output per human workday by mid-August, which this site covered at the time. Two months later, OpenAI changed the measure, moving from total output to success rates broken down by task length and task type, and the conclusion tightened accordingly: independent completion of long tasks is climbing fast, yet roughly three quarters of engineering work such as inference and serving still isn't finished without a human.

The report's limits are also clear: the data is internal, task difficulty is estimated by OpenAI itself, and there is no externally reproducible benchmark. If the next report discloses how completion is judged, the 46% figure will be easier to weigh.

Sources: OpenAI research acceleration report, AI Weekly, CocoLoop, Crypto Briefing; independent-completion rates by task length, success rates by task area and the median daily usage (converted at API list prices) were cross-checked.